This transcript is generated with the help of AI and is lightly edited for clarity.
//
REID:
Software is definitely not over. That’s hyperbole. But the key thing to track is it’s now much cheaper to make software. Even when you get to not just the—how does AI doing coding change all knowledge work? Even if you just say software becoming cheaper, what does that mean?
ARIA:
George Dantzig, in 1939, he was a mathematician in class. He came to class late, and there were two math problems on the board, and he wrote them down in his notebook and, you know, he went home and solved them and then came back to his teacher and said, like, so sorry I was late on the homework. And his professor was like, homework? Those were two of the famous unsolved math equations. He didn’t realize he shouldn’t persist. If we can reverse-engineer with AI a way to increase people’s persistence, how many more mathematicians would we have solving these unsolved problems?
REID:
AI is a cognitive exoskeleton. Part of exoskeleton, you realize you can do a lot more, like, oh my gosh, you can do these things. Wow, I can jump faster, I can run faster. And so I’m now more excited and interested in doing it.
//
ARIA:
Morning, Reid. Great to be here with you.
REID:
Good morning.
ARIA:
All right, let’s jump right in with something that Andrej Karpathy posted last week. And his sort of verdict was that the AI adoption gap has widened into a sharp funnel. So the way he sees it, there are 6 billion people on the planet who have barely touched LLMs. There are 1 to 2 billion people who have seen sort of a free-tier chatbot. There are another 20 million professionals that we’re talking about a lot, but specifically in math and code. And they’re watching agents complete in hours what used to take weeks. But then there are roughly 5,000 people with internal frontier access. And they’re seeing something else entirely. They’re seeing autonomous agents running cyber operations, advancing mathematics, and they are doing years of expert work.
And so his phrase for this last group was the vertigo of the curve going vertical. And it’s something felt by almost nobody. It’s felt just by these 5,000 people. And so is this just a normal way that technology moves and, you know, eventually the market will catch up and people will get to this sort of group of 5,000? Or is something structurally different about AI, and so this gap will take a lot longer to narrow, or maybe never narrow at all?
REID:
So narrowing gaps is kind of a little bit unclear. What I do think will happen is in terms of—you know, and Karpathy is always super smart in terms of how he looks at these things—that the whole funnel, like, people will go up these levels. It’s almost like he is kind of describing a pyramid, and what it is, actually, is people will be going through the pyramid. Now, whether or not a gap will be narrowing, I think, is an unclear question. What I do think is the thing is that most people have very little substantive experience—of, you know, of the 8 billion people—of what AI adds in their life.
And that’s one of the reasons why I was like, look, I think part of how we should, you know, help people understand how AI can be very humanity-positive, society-positive is, you know, do the medical assistant and the tutor and the legal assistant and the government services and all of these things in order to get out there. And I think it’s one of the things that any government that wants to help its people into the AI age should be doing with vigor and strength.
And you know, I’ve been recommending this to a number of governments over the years, and because they’re so slow at trying to figure out technological deployment for their citizenry, for their society, for, you know, kind of for industries, I’m just writing about it now so that people who actually want to go do something can talk to me about it. Now, that being said, I think that the most notable thing about Karpathy’s statement is not, is there a narrowing of this gap, but as much a couple of things. One, the fact that the folks who are working at these frontier models are seeing something that is very cutting-edge. And I think this is one of the things that, like, I’m very strongly in favor of OpenAI having released its mathematical proofs and so forth.
And I did see the various reactions of some mathematicians on this. And you know, we can spend a little bit of time debugging what is right and what is simply Ludditism in the mathematicians’ reactions. But it’s to show these internal models are doing really amazing things. Like, the things you see when you use ChatGPT or Claude or Gemini or Copilot from the outside, and from the kind of everyday users, are not actually, in fact, the cutting edge of what’s actually doable and possible now.
And that, I think, is one thing to really note, because—and part of that is like, you know, part of the lens of that was OpenAI saying, look, you know, we still have amongst the most leading models in the world. Look, we can do all this mathematics stuff. Let’s just release a bunch of the results. Which, you know, is not surprising. When you go, there were some mathematicians who were like, well, I was working on this for 10 years, and now AI has solved it. Well, it’s like, well, look, I was, you know, walking to work, and now there are cars. You know, it’s kind of like, okay, you know, that’s part of what happens, part of what the adoption is. And sure, you can say that the walking to work is part of the journey.
I get fitness, and I get time to think. And this driving car thing is kind of barbaric, but actually there are all kinds of utilities and new expanses of driving a car. Not to say you shouldn’t still walk sometimes. So that’s one comment from the Karpathy thing. The other is that there’s this kind of levels that go through it, and that most people’s exposure to AI—part of the reason why they have skepticism around data centers. Skepticism to say, you’re telling me that my job may be threatened, transformed, lost, et cetera, but what I’m seeing from my own use of AI is like an enhanced search. And that’s not a good trade for me. Right? Like, search works pretty well.
I’m happy to have enhanced search, but I don’t want to have a transformed job for that. And it’s actually—most people aren’t yet at the point where they’re seeing that their work is substantively changed in ways that go, oh my God, this is really amazing. And that’s why they also don’t have the exposure to go, well, where’s the place that I would still be working? Where’s the stuff that I would be doing? Like, if you have, you know, AI doing all of this work, where’s the place for me? Now, those of us who are using it still see huge roles for using AI and huge roles for us in decisioning and metacognition and a bunch of other things.
And so I think it’s also getting through these levels of the pyramid, going up to the major thing. And then, of course, the idea is to get as many people into the most productive and useful things for their lives, for their work, for their societies, for doing it. Now, back to the kind of closing of the gap. If the top is asymptoting, then the gap will close, but the top may still be increasing. So it may actually be that the gap is widening even as the water raises all of the boats, potentially a lot. So that was the reason why I was, like, uncertain on that particular part.
ARIA:
Let’s stick with power users today. So what are power users doing with AI? Actually, I was talking to my husband last week, and he told me about this story: a developer named Brandon Thomas, motivated by Adobe’s cancellation fees—he’s a power user—he used Claude Opus 5.5 to build seven free, open-source replacements for Adobe’s Creative Cloud. So Photoshop, Illustrator, Premiere, and four others. And he built it from scratch. There was no Adobe code touched, and so he called this a clean-room reimplementation. His comment was that software was over, especially because it’s not as if he was using discarded code or things from—he was just using Claude, and he recreated this thing that people do complain about a lot. And he said, here’s the open-source version.
And so this is new and different because something similar had been happening in gaming for a while, where AI had made decompilation cheap enough to be a hobby. And so modders were reverse-engineering binaries, porting console exclusives to PCs, they were reviving abandoned titles. And so all of these things are happening where people are essentially recreating software using AI. And so AI is really collapsing the cost of replicating what took these large teams years to build. And so is software over? Is it hyperbole? Like, what is going to be this next generation when someone can just vibe-code Adobe Photoshop in, you know, a few days, a few weeks, you know, a few afternoons?
REID:
Software is definitely not over. That’s hyperbole. But the key thing to track is it’s now much cheaper to make software. So previously it—and all kinds of things change from that. Like, it’s—even when you get to not just the—how does AI doing coding change all knowledge work? Which is an interesting—even if you just say software becoming cheaper, what does that mean? Well, for example, part of the reason why there was a difference between open-source projects and commercial projects was because, essentially, software was so expensive to build that only a limited set of things that were within the cost envelope could be built and deployed within an open-source environment. Because part of it is software doesn’t just live. It’s like you build once and you’re done.
It has to have security updates, it has to have functional updates, it has to have new OSes, it has to have new integrations, you know. And so it doesn’t—like, once a piece of software is deployed, it starts aging right away. And in the modern accelerated environment, it ages even quicker. And so that meant very limited room for open source. Now you make it much cheaper. It’s easier for individuals to do, it’s easier to create open source, it’s easier to upgrade it, it’s easier to maintain it. And so that means software is massively more plentiful. And that means that your moat of, we pay a thousand engineers to be maintaining a piece of software, and that’s why you buy our software, goes down a lot. And it doesn’t go down to zero, but it goes down a lot.
And that means the usual—because, you know, the margins, which caused the investment in the business and caused revenue and all the rest of the—become more challenging. And that’s part of the reason why people were talking about SaaS apocalypse now. Part of the reason I was, like, in the earlier thing, saying not SaaS apocalypse is I think it’s a tale of two software companies—you know, tale of two cities—which is, you know, those that are really adopting AI and figuring out what AI means now for how they’re providing their services, what their moats are. You know, for example, because software is plentiful, it could now mean that as opposed to one piece of software that is given to everybody, and everybody kind of mostly—like, everyone collectively dislikes it.
But it’s a little bit like Churchill on democracy. It’s the worst, except for everything else. It actually can be much more personalized given all the AI, and maybe, like, software becomes more of a platform in terms of how it operates, and then AI builds on top of it, and people still want the depth of the platform elements of it, and you can still have things. And so I think that, you know, that’s one possible path. I think there are a number of paths by which people will still be paying for software. My belief is you actually can even still get to, like, kind of good business margins on software. It’s just not what it was before, because software is a lot cheaper now.
You know, one part of it that I find to be very interesting about the, you know, the asymptoting to very low and still not zero for software—not just because, you know, someone who is a really good developer might, you know, create a whole bunch of open-source stuff and then it might be out there, but actually how you deploy and package and trust an open-source package. Now there’s going to be 40 open-source versions of, you know, Creative Cloud. And so, like, which ones do you use? They’ll have different versions and all the rest. Like, there’s just that. But it’s also, of course, a question of what happens now that software is so easily constructible. You know, does everyone have their own software running the house, for example?
I mean, there are just all these different expansions—that was meant to be deliberately somewhat trivial, but, like, there’s this range of possibilities. And that’s one of the things that’s been exciting me about this. And it’s also one of the things that suggests that, you know, part of what people frequently think entrepreneurship is, is, I have a eureka, I have a new product idea. And, you know, that’s great and important, but then, obviously, there’s eureka, I have a new way of doing product distribution, go-to-market, which is frequently more important, especially when there are plentiful products, which is, I think, you know, part of what, you know, kind of the internet revolution and the mobile revolution brought up. And then there’s the eureka, I have new business model ideas, which is the most rare.
But when it does hit, it’s huge, à la Google AdWords, et cetera, et cetera. And so I think this software going to, you know, kind of asymptoting to extremely cheap, I think, creates a wide range of additional—you know, kind of not just software products and ranges and new what it could be, but new go-to-market and new business models. And it’s across this whole spectrum that I think is really interesting. So with software is over, the thing that’s specifically wrong about that is, actually, software is amplified in importance and is changing across all these things. It’s just that the software game has, like, massively transformed. And that’s what I think is most interesting about creating the future.
ARIA:
And what do you think this does to IP? Like, do IP laws just become unenforceable? Do people try to enforce things even more aggressively because of where we’re at? Like, how does that change things?
REID:
So IP law is a human construction in order to create the incentive to create investment in IP that ultimately benefits, you know, humanity and society. And, you know, the simple definition, which I think will probably need to change, is we essentially give you a monopoly on something as a first constructor. And, you know, that’s, I think, still the right thing to do in pharma, because it’s so expensive and so important to do—you know, part of the reason why, you know, Manas AI. And, you know, it’s creating, you know, kind of monopoly drugs, you know, kind of as a way it’s operating. And so I think there are still some places, but I think there’s, like—it’d be interesting to think about, like, how would you, from first principles, reinvent, you know, IP law?
Now, we’ve always known that software IP law is amongst the most bizarre. It’s like, well, I have an idea that you can click on something, and with one click you can buy it, and that should be patentable. And you’re like, what? Like, totally, you know—because it wasn’t like it cost a billion dollars of risky investment in order to do this, which is part of what you’re trying to incent in order to make sure it happens. And it doesn’t mean that there shouldn’t be any software IP. There’s definitely—obviously, you know, one of the places where there’s also IP is brand and other kinds of things, which are important, because it’s like my identity. I’m building in trust and reliability and all of these features.
So I hedgehog my brand. But then the question of, like, well, maybe certain kinds of intellectual property construction are, you know, rare and risky and tricky enough that we need to kind of award some protection to doing it. And so—but I think that what it does definitely mean is that the current IP laws around software, which were already pretty mostly, you know, kind of frail and broken, are now even more so, from a—like, I’m trying to patent the action of the software, right, versus other things. And it’s unclear to me which parts of the actions of the software should actually, in fact, be patentable in a new thing. I haven’t really given it much thought, partially because we’re terrible at rethinking from first principles, like, you know, what should IP law, generally speaking, be in this case?
And it ends up mostly becoming a political fight where, you know, like, for example, software companies say, of course one-click for purchase should be patentable, because that adds value to us. And, you know, this is a little bit like—you know, one of the things that I was learning about recently is, for a number of years—you know, kind of started in Obama and wiped out by DOGE—there was, you know, the IRS sending the tax return to people, saying, hey, this is a tax return of all the data we have on you. Tell us if there’s anything you want to change or anything that’s incorrect here. But it means that you don’t have to go hire, you know, H&R Block or TurboTax or all the rest in order to submit it.
And it’s much easier, and totally—and, you know, easier for the person, easier for the government, easier all around. And, you know, DOGE got rid of this, maybe because of industry pressure, and so therefore it’s costing the entire society and all these people more. You want to have that kind of function and feature, and that’s part of the reason why you have to be much more careful about which things you’re actually creating IP protection around, because that’s the kind of—you know, kind of free to all society—is the kind of thing you want.
ARIA:
Yeah, I mean, it’s interesting when you bring up brand as it relates to IP, because I think some people could argue that brand’s going to matter less, because then you can vibe-code something over the weekend—who cares about brand? But I think other people, and I’m probably in this camp for the moment, are going to argue that brand matters more, because the only thing that’s going to matter is trust. And so when you’re using AI or when you—what to do, sort of the trust of that brand is going to be huge for security, for fidelity, for sort of all the things that you want. And so I think people will aggressively protect their IP as it relates to brand. But let’s go back to the cognitive space. You, again, mentioned tutor.
And I think a lot of the criticism that people have about AI, they talk about cognitive offloading. They say, oh, you know, you’re not going to remember the things that you need to know. And people always say, well, you know, the calculator was just like that too. You don’t need to, you know, multiply two-digit numbers in your head anymore, because you have a calculator in front of you. But I want to ask about a different kind of cognitive debt. So recently a randomized trial of 1,200 people—and this was out of CMU, Oxford, MIT, and UCLA—they found that just 10 minutes of AI use eroded people’s persistence. So once the AI was removed, people abandoned hard problems twice as often. And the authors called this cognitive debt.
And so this is essentially, like, grit, persistence. It’s like, I am going to try to solve that problem as hard as I can. Well, now that I’ve used AI, I am so used to sort of getting a quick fix or a quick answer, or this is what people could posit, that I’m going to quit more quickly. And a lot of people have been talking about this. But this study, presented at COLM this week, is the first to establish causation, not just correlation. And so this is, if you use AI, you will be less likely to be persistent. And so AI clearly makes people more capable in so many ways, but perhaps it also makes you less persistent. Like, what do you think?
You know, obviously we need a lot more studies, but what do you think of that trade-off? Does that concern you, that you could be more capable but sort of have less grit?
REID:
Well, as a macro—and my answer will completely not surprise you or probably many of our listeners—which is, look, there might be a challenge here. And if there is a challenge, it’s just a question of how we shape AI. So part of the reason why it’s worth doing this is to say, well, actually, in fact, we should shape AI to be increasing the capability of persistence. So, like, for example, you know, how do I learn things, you know, in kind of better ways? And so then I get it that they were contrasting the, you know, I’m given a hard problem without AI, and will I stay on it? And if I have a tool, it’s kind of like—like, for example, when I only had a bicycle, I might bicycle an hour into town.
And now that I have a car, maybe I’m not going to bicycle so much, because I’m like—I’m either like—and if I don’t have my car, I’m just not going to go, because I’ll wait for the car or we carry fixed or something else. And that’s not necessarily a bad function. But what you do want is you want persistence. All I it says so, like, for example, I have an amazing tool, the tool’s not there, I wait to have the amazing tool to work on it. That’s not a problem of persistence. A problem of persistence is I just am not really exercising my brain. I’m not trying to get better, I’m not trying to do it.
And so you want that persistence in the use of the tool, and generally, to be kept high. And I think AI can shape it. The other interesting question about the abandoning hard problems is part of what intelligence is, is making a, you know, kind of—and this is where prediction is important—is like, well, do I think I can solve this problem? So, for example, you could spend 10 years working on a hard problem and getting nowhere. And obviously maybe there was a chance you could get somewhere, and you should have spent it, and it was just an acceptable risk. But part of what we learn in wisdom and in judgment is, well, which problem should we actually, in fact, kind of persist on? Which should we put down and come back to later?
Which should we consult someone else on? Which do we say, well, that one’s probably just beyond me or beyond us, and I’ll see if someone else gets there? And I think that the—you know, and that kind of gets back to, look, AI is a cognitive exoskeleton. And actually, in fact, having that exoskeleton is actually, in fact, generally really good for us. And part of exoskeleton, you realize you can do a lot more, and part of it is to, like, you know—what I think, you know, mine and yours and all of our team have gotten into is like, oh my gosh, you can do these things. Wow, I can jump faster, I can run faster. I can, you know—mentally. And so I’m now more excited and interested in doing it.
It’s part of the thing about, like, AI being a really good tutor: I now want to learn a bunch of things that I otherwise wouldn’t have, because learning that thing would have been too hard to do, because it’s like, oh, you just buy a book and you work your way through the book, and you have enough time, and you don’t really have someone to talk to about it, and da, da, da, da. And all of a sudden it’s, you know, personalized, and you have an infinitely patient tutor to talk to about it. And so it’s not that I don’t think there’s anything—like, they may well have discovered a real issue. Oh, and then the last point for this is—
It’s a little bit like saying, hey, a person’s never seen a TV, now they’ve seen a TV, and they’re a little bit mesmerized by it, right? And it’s like, okay, so that’s a known thing. And the question is, obviously, as people grow up with it, what is their persistence going to be? And that’s the most important thing. And keeping persistence high mentally—learning—is a good feature, and that’s a feature we should be paying attention to. How do we shape the AI to make it happen?
ARIA:
It’s funny. This is reminding me—and I forget why I heard it recently—it was a UC Berkeley guy, I think, George Dantzig, in 1939. He was a mathematician in class. And I mean, it’s apocryphal, but it’s so fantastic. He came to class late, and there were two math problems on the board, and he wrote them down in his notebook, and, you know, he went home and solved them and then came back to his teacher and said, like, ugh, so sorry I was late on the homework. You know, these seemed harder than usual. And his professor was like, homework? Those were two of the famous unsolved math equations that I just wrote on the board. You solved two of them. Because he didn’t realize he shouldn’t persist.
He just thought this was homework. Of course, this is expected of me. And so if, you know, people say, like, if we can reverse-engineer with AI a way to increase people’s persistence, like, how, you know, how many more mathematicians would we have solving these unsolved problems? So to your point, we just need to be able to tune it so that we can get sort of the persistence and grit that is the optimal amount. Because there can also be too much. We know people who run through walls in a bad way. So—
REID:
Yeah, or keep trying to run through walls and never get through it.
ARIA:
Right. And never actually run through them. Reid, always a pleasure. Thank you so much.
REID:
Yep, likewise. Great to see you.
Possible is produced by Palette Media. It’s hosted by Aria Finger and me, Reid Hoffman. Our showrunner is Shaun Young. Possible is produced by Thanasi Dilos, Katie Sanders, Spencer Strasmore, Yimu Xiu, Aman Suri, Danny Garrison, Trent Barboza, and Tafadzwa Nemarundwe.
ARIA:
Special thanks to Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, Parth Patil, and Ben Relles.